diffbot VS tensorflow_cpp

I personally went through the hardest way possible and built my robots from scratch and it took 11 years and 3 degrees (BCs, MCs, PhD), but you don't need to go the same way. You can find some kit (for example https://github.com/ros-mobile-robots/diffbot) and start learning linux, bash, C++, Python, ROS, vehicle dynamics, computer vision, soldering and mechanical skills trying to keep this thing alive. It might sound too hard or even impossible, but solving the actual problems of your tiny robot and watching how it's making its first steps would motivate you to push forward. Also you will anyway learn a lot of skills that you will help you in the future.

Looks like tfcompile still exists, but I have never used it myself. Doesn't look to be much better documented than the C++ API itself. The full C++ API of course gives you way more flexibility and doesn't involve this special process of compiling a specific model. We have been pretty happy with our additional wrapper library tensorflow_cpp, allowing us to easily load arbitrary frozen graphs and saved models for inference.